Enhancing groundwater level prediction accuracy at a daily scale through combined machine learning and physics-based modeling

Yongding River alluvial proluvial fan, a part of North China Plain with intense groundwater withdrawals. The objective of this study is to enhance the accuracy of groundwater level (GWL) prediction at a daily scale by combining short-term memory (LSTM) and physics-based (PB) models. Two types of LSTM models were developed: LSTM-H-ORI, which predicts GWL, and LSTM-dH-ORI, which predicts simulated errors of the PB model. These models were subsequently refined into LSTM-H-IMP and LSTM-dH-IMP by incorporating the outputs from the PB model. To investigate the influence of groundwater flow processes, LSTM-H-IMP was further enhanced by considering the time lag properties of precipitation, referred to as LSTM-H-IMP-LAG. Significant improvements were found for LSTM-H-IMP when the performance of LSTM-H-ORI was poor (Nash–Sutcliffe Efficiency Coefficient, NSE<0) or the performance of PB model was not bad (NSE≥0). And the prediction accuracy was improved for over 67% of wells in this case. When the performance of PB model was medium and poor (NSE≤0.6), the improvement of LSTM-dH-IMP was more effective, leading to a prediction accuracy enhancement for over 77% of wells. Additionally, LSTM-H-IMP-LAG exhibited further improvement, with an average NSE increase of 0.1. This study provides scientific methods for accurate prediction of GWL at a daily time scale and the combined application of LSTM and PB models.